jk200201/qwen2.5-coder-7b-bird-cot

jk200201/qwen2.5-coder-7b-bird-cot: Advanced Text-to-SQL

This model is a 7.6 billion parameter text-to-SQL solution, fine-tuned from the Qwen/Qwen2.5-Coder-7B-Instruct base. It excels at generating SQLite queries by employing a chain-of-thought (CoT) reasoning process over database schemas.

Key Capabilities and Performance

• Execution-Verified CoT Distillation: The model was trained by distilling execution-verified chain-of-thought solutions from a powerful teacher model (Qwen3-Coder-480B), focusing on correct reasoning steps rather than just SQL answers.
• High Accuracy on BIRD Benchmark: Achieves 52.1% greedy execution result accuracy on the BIRD dev benchmark, a challenging dataset with realistic and complex schemas. With self-consistency (K=8), its accuracy rises to 58.5%, performing comparably to much larger models like DeepSeek V4-Pro (1.6T parameters) at a significantly reduced size and computational cost.
• Step-by-Step Reasoning: Designed to first reason about the schema and question, then output the final SQL query within a fenced code block.

When to Use This Model

• Text-to-SQL Applications: Ideal for converting natural language questions into SQLite queries, particularly for analytical tasks over complex, realistic database schemas.
• Resource-Constrained Environments: Offers high performance for its size, making it suitable for local deployment where larger models are impractical.
• Enhanced Accuracy with Self-Consistency: For critical applications, its self-consistency feature can be leveraged to boost accuracy, albeit with increased inference cost.

Limitations

• Primarily tuned for BIRD-style analytic SQL and SQLite dialect; may require adaptation for unusual dialects or domains.
• The highest accuracy (58.5%) requires K=8 self-consistency, which involves approximately 8x inference compared to the single-shot greedy accuracy of 52.1%.